Every time publisher traffic data gets worse, the same reassurance appears: the clicks you're losing were never worth much. People who bounced immediately. People who wanted a quick fact and left. AI Overviews are filtering out low-intent traffic, so the visits you keep are higher quality.
It's a comforting story, and until recently there was no clean way to test it. Correlational studies could show clicks falling alongside AI Overview rollout, but couldn't isolate cause, and couldn't say anything about what the surviving clicks looked like.
Saharsh Agarwal and Ananya Sen ran a randomised field experiment instead. They recruited 1,065 desktop users, ran a custom Chrome extension between 7 January and 10 February 2026, and randomised whether AI Overviews appeared. The working paper went up on SSRN on 3 April 2026 and was last revised on 8 July. The result is the clearest evidence we have, and it does not support the reassurance.
What they found
Two headline numbers, and the second is the one that matters.
Organic clicks fell 39.8% when an AI Overview was displayed, with zero-click searches up 34.5%. This comes from a randomised design rather than a before-and-after comparison, which is why it's worth more than the pile of correlational estimates that preceded it: the attribution is actually defensible.
Click quality showed no measurable difference. Across every metric they tested, the visits that survived an AI Overview looked statistically indistinguishable from the visits in the control condition:
| Metric | With AI Overview | Without AI Overview |
|---|---|---|
| Returned to search results | ~40% | ~40% |
| Visits ending under 10 seconds | ~18% | ~18% |
| Time on destination site | Statistically indistinguishable | |
The traffic that disappeared was not disproportionately made up of bouncers. It was a broadly representative slice of what you were getting before. You lost 39.8% of your clicks, and what you lost looked much like what you kept.
Why the distinction matters commercially
If the lost clicks had been low quality, a 39.8% traffic decline would translate to a much smaller revenue decline, and the correct response would be to do very little. Ride it out. Focus on the good traffic.
If the lost clicks were representative, a 39.8% traffic decline translates to roughly a 39.8% decline in whatever that traffic was producing. That's not a measurement artefact to be explained away in a board meeting. It's a straight hit to the channel that needs replacing from somewhere.
The gap between those two interpretations is the difference between a monitoring task and a strategy change. A lot of people have spent the past year on the wrong side of it because the reassuring version was never challenged with clean data.
Where the damage actually lands
The effect is not evenly spread, and this is the practical part.
It concentrates on informational queries. AI Overviews triggered, or were intended to trigger, on 53% of informational searches, against 15% of navigational and 6% of transactional ones. Navigational and transactional queries showed no statistically significant impact. If someone is searching your brand name or looking to buy, the answer layer isn't intercepting them.
Top-ranked results lost the most. When a top-of-page Overview was removed, the top three results absorbed the bulk of the returning clicks, with position one nearly doubling. The sites hit hardest are the ones that were winning. If you built a content programme around ranking first for high-volume informational terms, you are the exposed case, not the safe one.
The market-level picture is consistent, if messier than the headlines suggest. Publisher traffic declines are real but wildly dispersed depending on who's measuring and what they're measuring. Chartbeat put search traffic down around 33–34% globally and 38% in the US in the year to late 2025, and roughly 17% across Europe. Similarweb data across the top 40,000 US sites showed only about 2.5%. Both can be true: the damage is concentrated in news and informational publishing rather than spread evenly across the web.
Which is worth saying plainly, because the round numbers that circulate in decks ("publisher traffic is down 25%") usually turn out to be one dataset's figure for one segment over one period, quoted without any of that. Find out where your segment sits rather than adopting someone else's.
What this changes about your content plan
Not everything. But the following stop being optional.
Audit your traffic by query intent, not by page. Split your organic sessions into informational, navigational and transactional. The informational share is your exposure figure. If it's 70% of your organic traffic, you have a structural problem regardless of how your rankings look this week.
Stop treating ranking first as the safe position. For informational terms, position one is now the position most likely to be intercepted. That doesn't mean abandoning it; it means not building the forecast on it.
Move value behind the click. The pages that hold up are the ones an AI summary structurally cannot replace: calculators, tools, templates, proprietary datasets, anything interactive, anything requiring an account. A summary can reproduce your explanation. It cannot reproduce your tool. The complementary move is making sure the summary at least credits you when it does answer without you, which is a question of how small sites earn citations rather than of ranking.
Measure both layers. Search Console's Search Generative AI reports, live since 3 June 2026, give you impressions in Google's AI surfaces. GA4's AI Assistants channel, live since around 13 May 2026, gives you third-party assistant referrals, and tracking AI traffic in GA4 walks through the property configuration. Note that it deliberately excludes AI Overviews and AI Mode, so the two aren't halves of one funnel. Neither is complete alone. Together they tell you whether the answer layer is a growing channel for you or just a growing tax, and here's the method for joining them.
Don't reach for the opt-out. Google is rolling out a Search Console toggle that blocks your content from AI Overviews, AI Mode and AI Overviews in Discover, UK first, under CMA order. It explicitly isn't a ranking signal for search outside those features, so it's technically low-risk. Two things to weigh anyway. It doesn't cover the Gemini app. And for a small site, removing yourself from the answer layer means removing yourself from the one surface where you can appear alongside incumbents who outrank you everywhere else. Publishers with brand traffic to protect may reason differently. Most sites shouldn't.
The honest counterpoint
Three things worth holding alongside this.
The paper is a working paper, posted April, revised in July, not through peer review. The randomised design is a real methodological step up on what came before, but it isn't the last word.
The sample is 1,065 desktop users over five weeks. That's a legitimate size for this kind of design, and it is still one panel, one device class, one window.
And Google isn't necessarily arguing in bad faith. It's entirely plausible that internally-measured "quality" means something different from bounce rate and dwell time: satisfaction signals we can't see, on queries that never generate a click in either condition. That's not the same as being right about your revenue, but it's not the same as lying either.
The point
The reassurance was that you were losing junk. The best available evidence says you were losing a representative sample, and that the loss concentrates precisely on informational content ranking at the top: the exact profile most content marketing programmes were built to produce.
Plan for it as a real reduction in a real channel, work out how much of your traffic sits in the exposed category, and start building things a summary can't reproduce.
Sources
- Agarwal, S. & Sen, A., randomised field experiment on AI Overviews, SSRN abstract 6513059 (posted 3 April 2026, revised 8 July 2026)
- Search Engine Journal, "Google AI Overviews Study Finds Lost Clicks Weren't Lower Quality"
- PPC Land, "Researchers find Google AI Overviews cut publisher clicks 39.8%"
- Chartbeat data via Axios and Press Gazette, publisher search traffic decline (2025–2026)
- Similarweb / Graphite data via Press Gazette, top 40,000 US sites
- Nieman Lab, on publishers considering opting out of Google (July 2026)